Computational & AI · Computational & AI archive
Computational & AI — September 2026
2026-09-30 1 article
PLAC Dataset Tests Whether MM/GBSA Can Predict How PTMs Shift Ligand Binding
- The authors curated PLAC, an experimentally validated set of ligand activity changes caused by post-translational modifications; affecting PTMs tend to sit closer to the ligand than neutral ones.
- They benchmarked end-point MM/GBSA protocols across MD lengths and internal dielectric constants.
- Longer MD with a higher dielectric gave the best correlation with experiment, while a lower dielectric classified enhance/weaken/neutral better.
Nobody builds SAR around a PTM state, yet it can quietly explain why a potent binder fails in cells. A curated benchmark is worth more to me than yet another scoring tweak, because we finally have something to test against. The dielectric trade-off between ranking and classification is a useful warning: pick the protocol for the question you are actually asking.
2026-09-27 1 article
AI Pocket-Based Design Yields a Potent STAT3 Inhibitor for Colorectal Cancer
- A pocket-conditioned, autoregressive generative AI model was used to design candidate molecules directly within the STAT3 binding pocket, rather than screening pre-existing compound libraries.
- This structure-based generative approach produced a potent small-molecule STAT3 inhibitor that advanced to biological evaluation.
- The inhibitor showed activity against colorectal cancer, where STAT3 hyperactivation is a recognized driver of tumor growth and survival.
STAT3's SH2 dimerization pocket is shallow and dynamic — the reason most inhibitor programs (niclosamide analogs, peptidomimetics) never reached clinic. Generating atoms directly inside that pocket geometry, rather than rescoring a static virtual library, is the real methodological jump here. Still, "potent" on paper means little until we see SH2 selectivity over STAT1/5 and a synthesizable route — that's where de novo hits usually stall.
2026-09-24 2 articles
Two-Stage Generation Dynamics Enable Training-Free AI Drug Design
- Bayesian flow networks design molecules for protein pockets by jointly generating atom types and 3D coordinates.
- The two channels behave differently: atom types shift abruptly at a defined point, while coordinates sharpen gradually, a pattern consistent across model variants.
- Timing refinement to these transition points yields a training-free method matching baseline docking performance on a standard benchmark, with no extra predictor needed.
What earns this a second look is the lack of per-target retraining — that's the step that kills most generative tools before they reach my bench, so a training-free trick timed to a phase transition is worth trying. But matching baseline docking, not beating it, is a floor, not a win — and CrossDocked2020 scores say nothing about synthesizability or real binding.
Hierarchical Phase Transitions in Bayesian Flow Networks Enable Training-Free Molecular Design ↗
Rule-Based SMILES Editing Tool Aids Fragment-to-Lead Optimization
- A new method systematically edits SMILES strings—swapping atom types, resizing or opening rings, and grafting fragments—to generate analog ideas for fragment-to-lead optimization.
- First implemented as a KNIME workflow tested on fragments from multiple programs, it's now also available as open-source Python code paired with AutoDock Vina docking.
- This non-learned, rule-based approach complements AI/ML generative design, offering a practical option when training data for a target is scarce.
Honestly this is what I'd do by hand anyway—ring contraction, bioisostere swaps, fragment grafting—just automated and exhaustive instead of me sketching ten ideas on a whiteboard. For a scarce-data fragment series, I'd trust this over a generative model I can't sanity-check. Free KNIME/Python beats begging IT for another SaaS license too.
Fragment Optimization─Simple SMILES String Manipulation Strategies to Optimize a Hit or Fragment ↗
2026-09-23 2 articles
Charting a Path Through Billion-Compound Virtual Screening
- Commentary piece examining the rapid growth of make-on-demand virtual compound libraries, now spanning tens of billions of molecules, and the practical challenges this scale poses for computational screening.
- Discusses how researchers prioritize and search these vast chemical spaces efficiently, weighing computational cost against the risk of overlooking promising hits.
- Considers how such large-scale screening approaches fit into early-stage hit-finding workflows in drug discovery.
Scale for its own sake buys nothing if the scoring function can't discriminate at that volume — past a few hundred million compounds, enrichment from most current physics or ML scorers plateaus fast. This piece is refreshingly honest that the triage problem isn't solved; given how thin the specifics are, that caution is the most I'd take from it right now.
Docking Showdown: Physics-Based, Co-Folding, and Hybrid Models Face Off
- Compares physics-based docking, deep-learning co-folding, and hybrid rescoring methods on how well they generalize to unfamiliar protein targets and chemical scaffolds.
- Builds on the recent Runs-n-Poses benchmark, which found that classical physics-based docking held an edge over several co-folding tools on truly novel targets.
- Assesses pose accuracy and ranking performance to offer practical guidance on choosing a docking strategy based on target novelty.
Physics-based docking still beating co-folding tools on genuinely novel targets is the finding I keep coming back to — these deep-learning models are pattern-matching PDB coverage, not solving folding physics, and it shows the moment a target has no close analog. Until a co-folding tool proves itself on a zero-analog case, I'm running both and trusting the classical scorer as the tiebreaker.
Multiparadigm Benchmark of Molecular Docking: From Physics to Co-Folding and Hybrid Models ↗
2026-09-22 3 articles
Machine Learning and Docking Methods Show Uneven Generalization on DNA-Encoded Library Screens
- The study benchmarks machine learning, physics-based docking, and cofolding methods for predicting protein-ligand binding using DNA-encoded library (DEL) screening data.
- Methods compared include state-of-the-art ML models alongside physics-based docking tools (Schrodinger Glide, Rosetta GALigandDock) and the cofolding method Boltz-2.
- Ground-truth binding data came from ultralarge DEL selection experiments across three biologically diverse protein targets and multiple DEL synthesis formats.
- The work follows a NeurIPS competition in which even top-performing ML models trained on DEL data failed to generalize to out-of-distribution (OOD) chemical space.
- ML models performed well in-distribution, but out-of-distribution hit discrimination depended on the specific target and ligand context, with no single method consistently dominating across cases.
- The authors conclude that benchmark performance alone cannot predict OOD generalization and release an open-source package, DEL-iver, for evaluating binding-prediction methods against DEL screening data.
This is the paper you send to anyone pushing to swap your docking rescore step for a single ML model. "No method dominates OOD" is the whole ballgame: your real triage work is always on a scaffold the model hasn't seen, and a benchmark AUC built on in-distribution DEL data tells you almost nothing about how a model will rank the novel matter you bring in next week. I don't read this as an argument against ML - Boltz-2 and friends are genuinely useful, fast filters - but as a case for running physics-based docking/cofolding in parallel, not as a fallback you only reach for when ML looks shaky. Target- and ligand-context dependence also means there's no one 'best' pipeline to standardize on company-wide; you validate per target, per chemotype, every time. DEL-iver going open-source is the practically useful bit - a shared OOD benchmark beats trusting a vendor's Tanimoto-flattered internal numbers.
AI-Driven Lead Optimization Delivers 11B-6, a More Potent HBV Capsid Modulator
- Investigators describe a multi-objective lead-optimization campaign built on the sulfamoylbenzamide (SBA) chemotype exemplified by NVR 3-778, an earlier clinical-stage HBV capsid assembly modulator (CAM).
- An AI-guided computational workflow was integrated with iterative synthesis to simultaneously optimize antiviral potency alongside physicochemical and ADME/drug-like properties across the series.
- The campaign identified 11B-6 as the optimized lead, reported to combine stronger anti-HBV potency with improved drug-like characteristics relative to earlier series analogues.
- Mechanistically, 11B-6 functions as a capsid assembly modulator, engaging the HBV core (capsid) protein to misdirect nucleocapsid assembly and interfere with viral replication.
- The work positions 11B-6 as a candidate for further preclinical evaluation within the sulfamoylbenzamide CAM class for chronic hepatitis B treatment.
What actually catches my eye isn't the potency bump — it's that AI-guided optimization was wired into the synthesis loop, not run as a pre-filter before chemists took over. That's the difference between a screening deck and a real DMTA cycle. Going back to the SBA/NVR 3-778 chemotype rather than hunting a new capsid-binding scaffold is the pragmatic call: the pocket and assembly-misdirection mechanism are already de-risked clinically, so this campaign is buying SAR knowledge instead of paying discovery tax twice. That said, 'improved drug-like properties' is a claim I want backed by real microsomal stability, plasma protein binding, and hERG numbers before I get excited — potency plus a nicer cLogP on paper doesn't guarantee 11B-6 clears whatever tolerability hurdle likely sank NVR 3-778. Worth watching, not yet worth celebrating.
New Computational Score Aims to Predict PROTAC Oral Absorption
- PROTACs (proteolysis-targeting chimeras) are large, polar, conformationally flexible molecules that fall outside classical Rule-of-5 space, making oral absorption difficult to predict from standard physicochemical filters alone.
- The paper introduces a fully computational scoring approach for oral absorption, intended to support rational, in silico PROTAC design earlier in the discovery process rather than relying mainly on measured properties.
- The score is built from computed molecular descriptors likely spanning size, polarity/lipophilicity balance, flexibility, and conformational behavior such as intramolecular folding — features previously linked to permeability in beyond-Rule-of-5 compounds.
- By combining these descriptors into a single predictive metric, the method aims to help chemists rank or triage PROTAC candidates for oral-absorption potential before synthesis.
- The authors position the score as a practical design-stage tool for guiding linker, warhead, and E3-ligand modifications toward improved oral bioavailability, complementing rather than replacing experimental permeability assays.
Any tool that lets me deprioritize a bad linker before I burn two weeks on synthesis and a Caco-2 run is welcome — PROTAC medchem cycles are brutal, and "just make it and test permeability" stops being an option once your library is 50+ analogs deep. That said, I've seen enough beyond-Ro5 permeability models oversell intramolecular H-bonding and chameleonic folding to stay skeptical until I see the validation set and how it handles my actual chemotypes — conformational behavior is exactly the thing computation still fakes worst. Where this actually helps is triage, not truth: rank candidates, flag the obvious dead-on-arrivals, then spend the wet-lab budget on the borderline cases. That's a real upgrade over eyeballing MW/PSA/cLogP cutoffs, which everyone in the PROTAC space already knows don't apply. If the descriptors genuinely capture folding propensity rather than just correlating with size, this could meaningfully shift where we spend linker SAR effort.
2026-09-21 1 article
All-Atom Receptor-Aware Deep Learning Model Advances De Novo Peptide Binder Design
- Researchers developed REAPS, a geometric graph neural network for designing peptide binder sequences that models the target receptor as a fully observable, all-atom structural context, rather than relying on backbone geometry alone as typical inverse-folding methods do.
- In head-to-head evaluations against the widely used inverse-folding model ProteinMPNN, REAPS showed improved sequence recovery, structural fidelity, and interface-level biophysical properties for both linear and macrocyclic peptide binders.
- The team paired REAPS with structural hallucination in a closed-loop de novo design pipeline to generate novel peptide candidates without requiring a known starting scaffold.
- Candidates produced through this pipeline were synthesized and tested against the neurokinin-3 receptor (NK3R), where they showed experimentally confirmed functional activity.
Sequence recovery and "improved structural fidelity" benchmarks against ProteinMPNN are nice, but what I actually care about is whether the closed-loop hallucination + REAPS pipeline produced binders without a template scaffold — that's the real test, and NK3R hits with confirmed functional activity clear that bar. All-atom receptor context should, in principle, capture side-chain packing and H-bond networks that backbone-only inverse folding misses, which matters a lot for macrocycle design where conformational preorganization is everything. Before I'd plug this into a campaign: what's the actual hit rate and binding affinity on those synthesized NK3R peptides, not just "functional activity"? And does performance hold on a GPCR-like target outside training distribution, or is this another tool that shines on its own benchmark set and stalls on my target. Worth piloting on a live target before I swap out MPNN wholesale.
2026-09-19 2 articles
Stress-Testing AI Co-Folding Screens: Lessons from a Keap1–Nrf2 Case Study
- Multi-stage virtual screening cascades that combine AI co-folding models with learned-affinity (ML) scoring are now standard practice in computational drug discovery.
- Each stage's output score is typically taken at face value and used to advance or discard candidates without independent validation of what it actually captures.
- This preprint systematically evaluates, stage by stage, whether cascade scores provide the information practitioners need to trust them, using the Keap1-Nrf2 protein-protein interaction as a case-study target.
- The authors characterize failure modes that arise at different stages of the cascade, offering a quality-control perspective on where and why these pipelines can mislead.
Every screening campaign now runs some flavor of co-fold-then-ML-score cascade, and honestly most of us take the ranked list at face value because checking each stage is a pain. That's the exact habit this preprint is poking at, using Keap1-Nrf2 — a notoriously flat, shallow PPI interface that's already brutal for docking and pharmacophore methods — as a stress test where scoring artifacts should show up fast. If the abstract's framing holds, the real message isn't "don't use these tools," it's "don't trust a single composite score without knowing what each stage is actually measuring" — pose confidence, affinity ranking, and physical plausibility are not the same thing, and conflating them is how false positives sneak to the top. Worth revisiting once the full methodology and failure-mode breakdown are out; details here are still thin.
New "Scope" Chemical Space Yields Ligand-Efficient Hits Across Three Targets
- "Scope" describes a mid-sized (10k-50k compound) chemical space positioned between minimal fragments and fully elaborated low-molecular-weight libraries, designed to maximize ligand efficiency and medicinal chemistry tractability.
- The authors introduce "Scope ring systems" (ring-X-ring) as a core organizing concept for building libraries and guiding functionalization toward sufficient potency.
- A 10k-compound Scope screening set — evaluated by SPR, native mass spectrometry, DSF, and biochemical assays — matches general diversity sets' primary hit rates even at higher screening concentrations, while showing superior ligand efficiency and higher hit rates after counter-screening.
- Case studies on trypanosomal CLK1 kinase, a 14-3-3/estrogen receptor-alpha molecular glue project, and SARS-CoV-2 main protease show the set yields diverse, tractable, ligand-efficient hits across varied binding sites.
The "sweet spot between fragments and full libraries" pitch isn't new — DEL, FBDD, and house libraries have all claimed it before. What's actually new in Scope is the ring-X-ring organizing logic: forcing vectors off two ring systems rather than one flat scaffold gives a built-in growth vector for SAR from hit one, the part fragment campaigns usually bolt on later. The number that matters to me isn't the hit rate, it's "higher hit rate after counter-screening" — that's fewer SPR artifacts and aggregators wasting my week. The real test is the 14-3-3/ERα glue case: PPI/glue targets chew up conventional libraries because there's no deep pocket to reward LE. If a rigid, moderately-sized scaffold set can seed a glue hit, that's the transferability claim worth watching, not the Mpro or CLK1 wins.
2026-09-18 3 articles
Benchmark Study Reveals Blind Spots in AI-Based Virtual Screening for Keap1-Nrf2
- A new benchmarking study examines a virtual-screening pipeline that chains structure-based co-folding models with learned-affinity scoring, an increasingly common setup for computational hit-finding.
- The cascade was applied to the Keap1-Nrf2 protein-protein interaction, a notoriously hard-to-drug target due to its flat, polar interface.
- Instead of trusting only the final ranked output, the authors assessed each stage separately against experimentally meaningful signal.
- They identify specific points where a stage's score diverges from that signal, a failure mode that can mislead downstream compound prioritization.
- The paper proposes per-stage control checks practitioners can run before relying on a cascade's final hit list, framing the work as a methodology and reliability study rather than a report of a new inhibitor.
Keap1-Nrf2 is exactly the flat, polar PPI where a co-folding model will happily spit out a confident-looking pose and the affinity scorer will rank it #1 for the wrong reasons — so a paper that pries the cascade apart stage-by-stage is more useful to me than another "we found a hit" claim. What I actually want out of this: a checklist to run on my own vendor or in-house pipeline before I let its top 20 dictate what goes on the synthesis queue. If the co-folding confidence and the affinity score are quietly diverging, that's exactly the failure mode that burns weeks on compounds that never had real signal behind them. No new chemical matter here, no binding mode to chase — just a reminder that for a target this treacherous, the ranked list is not the QC step. Filed under "read before ordering," not "read for the inhibitor."
Large-Scale Virtual Screen Delivers Potent New TAAR1 Agonists for Schizophrenia
- TAAR1 is a G protein-coupled receptor pursued as a non-dopaminergic drug target for schizophrenia and antipsychotic therapy.
- A structure-based virtual screen docked about 65 million candidate molecules against an active-state model of TAAR1.
- Of 55 top-ranked hits tested experimentally, 14 showed genuine TAAR1 agonist activity, ranging from mid-nanomolar to micromolar potency.
- Hit optimization then yielded agonists with nanomolar potency, and cryo-EM structures confirmed the binding poses the docking had predicted.
- Three optimized agonists showed high brain exposure and potency described as exceeding that of ulotaront, a clinical-stage TAAR1 antipsychotic candidate.
- In mice, all three normalized amphetamine-induced disruption of pre-pulse inhibition, a schizophrenia-relevant behavioral measure, without causing catalepsy, a common antipsychotic side effect.
A 25% experimental hit rate (14/55) out of 65 million docked molecules is the number that jumps out — that's an unusually clean confirmation rate for an active-state GPCR model, and it's real validation for structure-based virtual screening on TAAR1, not just a docking-score exercise. Better still, cryo-EM backs up the predicted binding poses, so the model isn't just getting lucky on affinity. The pharmacology is the real prize: brain-penetrant, potency beating ulotaront, and no catalepsy in the PPI assay — that's the non-dopaminergic promise actually paying off, not just a slogan. Ulotaront is now the bar every TAAR1 agonist has to clear, and these three cleared it. Worth watching whether the SAR generalizes beyond these three chemotypes.
Structure based discovery of antipsychotic-like TAAR1 agonists ↗
New Fragment-Based Docking Method Targets Long, Flexible Peptide Ligands
- Researchers built STELLAR, a new docking workflow aimed at a persistent problem: current tools poorly handle long, highly flexible biopolymer ligands, including peptides exceeding roughly ten subunits, because of their vast conformational freedom.
- STELLAR splits such ligands into smaller fragments, docks each one separately, then reassembles them into a full-length, physically realistic pose, using established software including GNINA, RDKit, and GROMACS for refinement.
- Tested against known peptide-protein complexes from the Propedia database and the Protein Data Bank, the workflow reproduced experimentally observed binding poses with high accuracy.
- Compared with other state-of-the-art docking tools, it runs mainly on CPUs, scales linearly with ligand length, executes faster, and includes a ranking metric suited to high-throughput screening of long-chain polymer-protein interactions.
Ten residues is roughly where my patience with standard docking runs out — past that, chains have so much conformational freedom that I end up either restraining everything by hand or giving up on the exercise. Fragment-then-reassemble is a sensible workaround, and leaning on GNINA/RDKit/GROMACS instead of a black-box neural net means I can actually inspect intermediate poses. CPU-only, linear scaling is the detail that matters most day-to-day: no GPU queue, no cloud bill, just run it on the workstation overnight for a stapled-peptide library. Most ML docking tools give me a score with no real confidence handle for ranking; a purpose-built ranking metric for long-chain ligands is genuinely missing elsewhere. That said, Propedia/PDB benchmarks are known complexes — reproducing a crystal pose isn't the same as correctly ranking novel binders against a target with no solved structure. Worth a pilot run, not a blind swap-in.
STELLAR: A Fragment-Based Docking Workflow for Highly Flexible Long-Chain Biopolymers ↗
2026-09-17 1 article
AI Screening Pipeline Flags Repositioned CASP4 Inhibitor Candidates for Alzheimer's
- CASP4, an inflammatory caspase linked to Alzheimer's disease, is the target of a new computational drug-repositioning pipeline built to screen DrugBank compounds for potential inhibitors.
- The workflow combined docking-based prioritization with ChemBERTa language-model embeddings and physicochemical descriptors to train machine-learning classifiers.
- A Random Forest classifier and regressor were used to identify likely active compounds and rank them by predicted potency.
- Downstream cheminformatics, PK-PD evaluation, and molecular docking narrowed the pool to five repositioning candidates.
- Molecular dynamics simulations and MM/PBSA free-energy analysis were then applied to assess binding stability, identifying two compounds (DB05316 and DB00519) with predicted binding compared favorably to the reference drug donepezil.
- The authors present the pipeline as a general strategy for prioritizing repurposed compounds against novel drug targets.
A five-stage funnel — docking, ChemBERTa embeddings, Random Forest, then MD/MM-PBSA — down to two compounds looks disciplined, but each stage has its own error rate, and stacking filters doesn't guarantee the errors cancel out. MM/PBSA rescoring of structurally unrelated ligands is notoriously unreliable for absolute ranking; chain enough filters and you can end up polishing false positives rather than eliminating them. Where this actually sits in the AI hit-finding wave: not novel chemistry, but faster triage of known-compound space, which is a real and growing niche. The genuine advantage is DrugBank sourcing — DB05316 and DB00519 already carry ADMET, safety, and formulation data, so there's no synthesis or fresh tox package standing between candidate and IND-enabling work. None of that substitutes for a binding assay. Until CASP4 activity is confirmed biochemically, this is a ranked hypothesis list, not a hit.
2026-09-16 1 article
ChemBERTa-Guided Virtual Screening Identifies Repurposable CASP4 Inhibitors for Alzheimer's Disease
- Ligand-based virtual screening pipeline mined DrugBank for CASP4 inhibitor candidates, combining docking-score prioritization with SMILES-derived ChemBERTa embeddings and physicochemical descriptors.
- A Random Forest classifier trained on the hybrid ChemBERTa/physicochemical feature set separated active from inactive compounds with high accuracy and enriched actives among top-ranked hits.
- A companion Random Forest regressor, trained on experimental pIC50 values, ranked candidates by predicted potency.
- Cheminformatics filtering, PK-PD evaluation, and molecular docking narrowed the field to five DrugBank-listed candidates.
- Molecular dynamics simulations and MM/PBSA free-energy analysis confirmed stable binding poses for the top candidates.
- Positions the workflow as a drug-repositioning route toward CASP4-targeted Alzheimer's disease therapy, a caspase linked to neuroinflammation and pyroptosis.
What actually sells this to me is the DrugBank sourcing, not the screening funnel itself: the five candidates that survive docking, ChemBERTa/RF classification, and MM/PBSA already come with known ADMET and need no de novo synthesis to reach an animal model — that's a tractability shortcut a genuine hit-finding campaign rarely gets. CASP4's pocket is the usual caspase catalytic Cys-His dyad, so the real question the MD trajectories should answer is whether these scaffolds actually engage that dyad or just sit in a peripheral exosite; binding energy alone won't tell me that. Selectivity is the open flag — caspase-1/3/4 share fold homology, and off-target caspase hits in a neuroinflammation context could cut the wrong way. Methodologically, ChemBERTa embeddings layered onto classical descriptors for RF triage is a sane upgrade over pure docking-score ranking, though I'd want held-out chemotype validation before trusting the pIC50 regressor outside its training distribution.
2026-09-14 1 article
AI Co-Folding Model Boltz-2 Speeds Discovery of Novel AKR1B10 Inhibitors
- AKR1B10 is an aldo-keto reductase enzyme overexpressed in several cancer types, making it an attractive but challenging drug target.
- Researchers used Boltz-2, an AI-based protein-ligand co-folding model, to computationally screen a large, chemically diverse compound library and prioritize a small, diverse shortlist of candidates for lab testing.
- Experimental testing of that shortlist produced an unusually high hit rate, with most of the tested compounds showing real inhibitory activity against AKR1B10.
- Several confirmed hits reached submicromolar potency and carried chemical scaffolds not previously reported as AKR1B10 inhibitors.
- The results support co-folding AI models as a practical, resource-efficient way to convert computational predictions into experimentally validated drug leads, reducing reliance on brute-force screening.
• Most of a small, deliberately diverse shortlist came back with real inhibition. That hit rate isn't HTS-style luck; it's a model that has actually learned something about the pocket, not the kind of docking score I'd normally discount. • Novel scaffolds are the real prize, but I want SAR before I get excited: is potency riding on one privileged fragment, or is there a real vector to optimize? Submicromolar out of the gate on unprecedented chemotypes usually means headroom, not a dead-end analog of known reductase binders. • AKR1B10 sits close to AKR1B1 (aldose reductase) — selectivity is the whole game here; nobody needs another dirty aldo-keto reductase pan-inhibitor with off-target baggage. • This is one target, one campaign. I'd want to see it replicated on something without such a cooperative pocket before I trust co-folding over ensemble docking/FEP as a default triage step. Promising, not proven — but the direction hit-finding is heading.
Efficient Discovery of Potent AKR1B10 Inhibitors Using Co-folding Model Boltz-2 ↗
2026-09-12 1 article
Multimodal AI Framework Targets Drug-Resistant Bacteria While Screening for Toxicity
- Antimicrobial resistance is narrowing treatment options for pathogens like Staphylococcus aureus and Neisseria gonorrhoeae, motivating new computational discovery tools.
- MAPViS, a multimodal deep-learning framework, fuses molecular graph, fingerprint, and descriptor/3D geometric features via gated fusion, pretrained on ~10 million molecules using hierarchical pseudolabel classification and masked graph contrastive learning.
- After fine-tuning, it jointly predicts antibacterial activity against S. aureus and N. gonorrhoeae plus cytotoxicity risk in three human cell lines (HepG2, HSkMC, IMR-90).
- On benchmark and external validation data, MAPViS showed strong early-enrichment performance for S. aureus, with all ten top-ranked candidates confirmed experimentally active, while generalization to N. gonorrhoeae was more modest.
- The authors describe MAPViS as a scalable tool for joint activity-toxicity modeling to support antibiotic virtual screening.
Ligand-based, not structure-based — no target structure in the loop, so I'd worry about applicability domain once you push past its ~10M-molecule pretraining chemotypes into a new target class. The joint activity/cytotox head is the real differentiator: folding HepG2/HSkMC/IMR-90 flags into the same ranked list saves the counterscreen step I'd otherwise run manually — genuine time saved, not just convenience. But the S. aureus (10/10 confirmed) vs. N. gonorrhoeae ("modest") split is the tell: performance tracks training-data density per pathogen, not real antibacterial chemotype intuition. Before trusting a ranked hit list on an unfamiliar target, I'd want per-pathogen calibration and confirmed true negatives, not just top-10 hit rates. Good as a cheap pre-HTS triage filter; not a substitute for structure-based design or real counterscreening.
2026-09-11 2 articles
Reversible-Work Metadynamics Delivers Fast Hit Triage — and Exposes a Memorization Blind Spot in AI Co-Folding
- Virtual screening often yields false positives from fast, approximate scoring functions, while accurate free-energy calculations to filter them remain too costly for routine high-throughput use.
- The authors introduce CTMD (c(t)-based metadynamics), which ranks binding stability using the nonequilibrium reversible-work estimator c(t) from short, independent metadynamics runs, without requiring converged binding free energies.
- Across diverse targets and chemotypes, CTMD shows robust early enrichment while staying fast, transferable, and largely tuning-free.
- Benchmarking AI co-folding, the authors find Boltz-2's enrichment tracks similarity to its training data and persists even after active-site modifications, pointing to memorization rather than true binding physics.
- CTMD is offered as an open-source middle ground between quick docking/co-folding and costly free-energy methods, cutting wasted discovery-campaign resources.
A triage tier between docking and full FEP is exactly what's been missing from my workflow. Standard docking scores are too noisy to prioritize synthesis, and running FEP on every virtual-screening hit is a non-starter for budget and turnaround — CTMD's short, independent metadynamics runs with the c(t) estimator could slot right into that gap and actually change which compounds we send to synthesis, without a cluster's worth of compute. Open-source and largely tuning-free means it's adoptable without a specialist babysitting it. The Boltz-2 finding is the one I'd flag to my team before anyone leans on AI co-folding for real decisions: enrichment tracking training-set similarity even after active-site mutation means it may be pattern-matching known chemotypes, not scoring genuine binding physics. That's a trap for exactly the novel scaffolds we're chasing.
Early-Enrichment Hit Discovery via Reversible-Work c(t) Estimation in Metadynamics (CTMD) ↗
Many Virtual Screening 'Improvements' May Not Be Real, Pre-Registered Study Finds
- The preprint asks a foundational methods question: given the size of a typical structure-based virtual screening benchmark, how large a difference in AUROC or enrichment between two protocols can actually be distinguished from statistical noise.
- Rather than assuming benchmark comparisons are inherently reliable, the authors first directly measure these resolution limits.
- They then pre-registered six candidate interventions to docking/scoring protocols and tested each against the measured resolution limits, rather than reporting post-hoc 'wins.'
- The pre-registration design is meant to guard against selective reporting, a known issue where small, benchmark-specific gains get presented as general improvements.
- This is a methodology/rigor study, not a new-inhibitor or target-specific paper; it addresses how virtual screening improvements should be evaluated.
- Practical implication: many published claims of improved virtual screening performance may fall within the noise floor of typical benchmarks, so hit-finding teams should treat single-benchmark improvement claims cautiously.
Every VS vendor pitch should have to clear this paper's bar. If a 'protocol improvement' can't beat the measured noise floor of AUROC/enrichment on a benchmark our size, it's not an improvement — it's a coin flip dressed up in a webinar slide. Six pre-registered interventions is a high bar; realistically most of us won't pre-register anything before a campaign, but I can at least ask collaborators to report effect sizes against a resolution limit, not just a headline AUROC bump. Sharpest implication for me: the co-folding hype (Boltz-2 and friends) needs the same treatment — if those models are just pattern-matching PDB-adjacent chemotypes, their 'wins' over classic docking/FEP on public benchmarks could be memorization, not generalization. I'm not swapping FEP for a black box on this evidence alone.